Body of Secrets: Anatomy of the Ultra-Secret National Security Agency (96 page)

Feeding
streams of intercepts from the worldwide listening posts to the analysts at NSA
is a special highly secure Sigint Communications System. First opened on the
eve of Pearl Harbor, the system carried over 25 million words a day by the
mid-1960s. Analysts using Harvest would then further process the encrypted
traffic.

Another
system bears critically important intelligence from an intercept operator at a
listening post in a distant part of the world straight to the president of the
United States at breakneck speed. The surprise launch by the Soviet Union of
Sputnik
in 1957 caused an earthquake within the intelligence community. At the
time, it took an average of 8 hours and 35 minutes for a message containing
critical intelligence to reach the White House. President Eisenhower demanded
that the time be reduced to minutes. At a National Security Council meeting on
August 27, 1958, attended by Eisenhower, CIA director Allen Dulles agreed that
"there was little purpose in developing critical intelligence overseas
unless we had the communications means to insure its rapid transmission to
Washington."

A month
later, in a meeting in the Oval Office with Eisenhower, Tordella proposed a
system known as CRITICOMM. After Tordella outlined the costs and benefits,
Eisenhower turned to the deputy secretary of defense and said, "Do
it." Within six months NSA was able to reduce transmission time from more
than 8 hours to 52 minutes. In another six months the agency was able to have a
CRITIC, or critical intelligence message, on Eisenhower's desk within a brief
thirteen minutes, regardless of where it had originated. Eventually the time
shrunk to between three and five minutes.

Finally, a
system codenamed Rye provided remote access to Harvest, thus permitting
analysts throughout NSA to access the main computer via several dozen distant
terminals. "RYE has made it possible for the Agency to locate many more
potentially exploitable cryptographic systems and 'bust' situations," said
one secret report at the time. "Many messages that would have taken hours
or days to read by hand methods, if indeed the process were feasible at all,
can now be 'set' and machine decrypted in a matter of minutes. . . . Decrypting
a large batch of messages in a solved system [is] also being routinely handled
by this system."

Few could
have foreseen Harvest's bright future when the machine was first built. Because
the complexity of the system baffled even many of the best analysts, it was
originally considered a white elephant. During employee tours, the huge, boxy
machine was pointed to and mocked. "It's beautiful, but it doesn't
work," officials would scoff. But once the machine was fully understood,
Harvest became so successful that it was used continuously for fourteen years.
The agency finally switched to a more advanced system only in 1976.

As
computers more and more became essential in both codemaking and codebreaking,
worries developed over the progress the Soviet Union was making in the field, especially
given its early lead in space exploration. In 1959 a top secret panel was
created to investigate where the United States stood in its computer race with
Russia. The results were encouraging. By then the U.S. government had about
3,000 computers, of which about 300 were high-performance machines valued at
more than $1 million each. Russia, however, had fewer than 400 computers, of
which only about 50 were large machines.

Although
for a time both countries attained comparable speed— the Soviet M-20 was about
as fast as the IBM 709—the United States had left Russian computer scientists
in the dust with the development of the transistor. Nevertheless, the secret
panel's report advised against overconfidence. "The Soviet Union could
achieve a computer production capability equivalent to that of the U.S. in 2—3
years, if they place the highest possible priority on the effort." But,
the report added, "There is no evidence that they intend to establish such
a priority." Nor, the report said, was the Soviet Union engaged in
anything equivalent to Project Lightning.

Following
Harvest, NSA's brain, like that of a human, was divided into right and left
hemispheres, codenamed Carillon and Lodestone. Carillon was at one time made up
of IBM 360s, and later of four enormous IBM 3033s linked together and attached
to three IBM 22,000-line-per-minute page printers.

Even more
powerful, however, was Lodestone. Dominating the center of a yellow-walled,
gold-carpeted hall of computers, front-end interfaces, and mass storage units,
was a decorative, 4½ -foot-wide, 6½ -foot-high semicircle of narrow gold and
deep green panels surrounded by a black vinyl-upholstered bench-type seat. It
appeared to be an ideal resting place for lunch or a mid-morning coffee break.
It was, however, the fastest, most powerful, and most expensive computer of its
time.

Built by
Cray Research at its plant in Chippewa Falls, Wisconsin, a town also known for
its Leinenkugel's beer and Chippewa Springs water, the $15 million CRAY-1 may
be the ultimate testimony to the old proposition that looks are deceiving.
Housed within what one wag once called "the world's most expensive love
seat" were more than 200,000 integrated circuits, each the size of a
thumbnail, 3,400 printed circuit boards, and 60 miles of wire. So compact was
the five-ton, seventy-square-foot unit that enough heat was generated per cubic
inch to reduce the machine to a molten mass in seconds had it not been for a
unique Freon cooling system using vertical aluminum-and-stainless-steel cooling
bars that lined the wall of the computer chassis.

The
supercomputer was the brainchild of Seymour Cray, a shy, enigmatic engineer who
rarely allowed interviews or pictures but was one of the most influential
figures in computer science. The founder of Cray Research, Inc., Cray "is
to supercomputers what Edison was to light bulbs," said
Time
in
1988, "or Bell to the telephone." When not in his laboratories, Cray
could likely be found deep in the earth beneath his Wisconsin home, slowly
tunneling toward the nearby woods. Eight feet high and four feet wide, the
tunnel was lined with four-by-four cedar boards. When a tree once crashed
through the roof of the tunnel, Cray turned the hole into a lookout with the
installation of a periscope.

To Cray,
the tunnel was both inspiration and recreation. "I work when I'm at
home," he once told a visiting scientist. "I work for three hours and
then I get stumped, and I'm not making progress. So I quit, and I go to work in
the tunnel. It takes me an hour or so to dig four inches and put in the
four-by-fours." Half kidding, Cray continued: "Now, as you can see,
I'm up in the Wisconsin woods, and there are elves in the woods. So when they
see me leave, they come into my office and solve all the problems I'm having.
Then I go back up and work some more." According to John Rollwagen, then
chairman of Cray Research, "The real work happens when Seymour is in the
tunnel."

Cray began
his career by building codebreaking machines in the 1950s with Engineering
Research Associates, then headed by future NSA research chief and deputy
director Howard Engstrom. Cray's dream was to build a number cruncher capable
of 150 to 200 million calculations per second. It would have between 20 and 100
times the capacity of then current general-purpose computers—the equivalent of
half a dozen IBM 370/195s.

In the
spring of 1976 the first CRAY-1 rolled out of the firm's production plant in
Chippewa Falls and directly into the basement of NSA. A second was quietly
delivered to NSA's secret think tank, the Communications Research Division of
the Institute for Defense Analysis at Princeton University.

The CRAY
had a random access semiconductor memory capable of transferring up to 320
million words per second, or the equivalent of about 2,500 300-page books; NSA
could not have been disappointed. And when it was hooked up to the computer's
specialized input-output subsystem, the machine could accommodate up to
forty-eight disk storage units, which could hold a total of almost 30 billion
words, each no farther away than 80 millionths of a second.

In a field
where time is measured in nanoseconds—billionths of a second—seven years is an
eternity. Thus it was with tremendous excitement that in June 1983 the agency
made space in its basement for a new arrival from Chippewa Falls, the CRAY
X-MP. Serial number 102 stamped on its side, the machine was the first X-MP to
be delivered to a customer; NSA thus had the most powerful computer in the
world at the time. The six-ton brain, which contained forty-five miles of wiring
and required a fifty-ton refrigeration unit to keep it cool, was revolutionary.
Rather than achieving its gains in speed simply by using a faster processor,
the X-MP used two processors, working in parallel. Two separate jobs could be
run at the same time, or one job could run on both processors. This capability
made the X-MP five times faster than even the most advanced CRAY-1, the
CRAY-IS/1000.

To NSA,
parallel processing was the wave of the future. Among the projects the agency
was closely involved with was the Butterfly processor, which linked 148
microprocessors. Developed by the Defense Advanced Research Projects Agency's
(DARPA's) Strategic Computing Program, Butterfly could have been scaled up to
combine 256 or 512 or even 1,000 linked processors. Future testing included
plans to link about 1 million processors.

The X-MP
arrived just in time. That same year NSA secretly put into operation an enormous
worldwide computer network codenamed Platform. The system tied together, into a
single cyber-web, listening posts belonging to NSA, GCHQ, and other Sigint
agencies around the world, with NSA as the central brain.

Two years
later, in 1985, NSA's basement complex became even more crowded with the
long-awaited arrival of the CRAY-2. With its bright red Naugahyde base and
transparent, blue-tinted towers of bubbling liquid coolant, Seymour Cray's
latest masterpiece looked more like bordello furniture than a super number
cruncher in a codebreaking factory. Nicknamed Bubbles, the $17.6 million
computer was almost human, with cool, bubbling Fluorinert, also used as an
artificial blood plasma, running through its system. The liquid was necessary
to keep the enormous heat generated by electrons flowing through the tightly
packed circuit boards from causing a meltdown.

The unit
of speed used in assessing supercomputers is the "flop,"
"floating point operations per second." Whereas it may take the
average person several minutes to calculate with a pencil the correct answer to
a single multiplication problem, such as 0.0572 x 8762639.8765, supercomputers
are measured by how many times per second they can solve such problems. If it
takes one second to come up with the answer, including where to place the
"floating" decimal point, then the computer is said to operate at one
flop per second. Bubbles, on the other hand, was able to perform at an
astonishing 1.2 gigaflops, or 1.2 billion mathematical calculations a second.
This made it up to twelve times faster than its predecessor and 40,000 to
50,000 times faster than a personal computer of that time.

By 1988
workers were laying wires and arranging power for still another new product
from the backwoods of Wisconsin, the CBAY Y-MP. So dense were the chips on the
new machine that engineers were now able to squeeze eight processors into a
space originally designed for only one. Working together, and under ideal
conditions, the processors were capable of performing between 2 billion and 4
billion operations a second.

In the mid
to late 1980s, the pace of supercomputer development was so fast that NSA
barely had enough time to boot up each new mega-machine before a newer one was
wheeled into its basement "flop house."

The race to
build the fastest supercomputer began to resemble a mainframe Grand Prix.
Sleek, shiny, and ever more powerful new machines were continuously zooming to
the starting line while engineers worked on ever more powerful and speedy
designs. Nobody wanted to be left in the dust. In September 1987, Steve Chen,
the Chinese-born computer superstar who lead the Cray Research design team on
the X-MP and Y-MP projects, left Cray after his machines became too expensive
and risky. He was quickly hired by IBM. "Five years from now,"
boasted an IBM executive, "we should be at 100 billion gigaflops. A
problem that takes three months to do now, we want to do in a day."

Off in the
shadows, the Sandia National Laboratory, in Albuquerque, was tweaking a chunky
little blue box. Three feet on a side and known as the Ncube, or hypercube, the
computer was "massively" parallel, with 1,024 processors, each as
powerful as a traditional minicomputer. In a test, Sandia asked the computer to
calculate the stresses inside a building beam supported only at one end. A
powerful minicomputer working twenty-four hours a day would have taken twenty
years to arrive at an answer, but the lightning-fast Ncube accomplished it in a
week.

At ETA, a
subsidiary of Control Data Corporation, a dark, bubble-topped box known as the
ETA 10 was unveiled. An eight-processor powerhouse, it used computer chips that
were smaller and denser than those used by Cray Research. Liquid nitrogen
carried away the excess heat. And by using only one circuit board, the engineers
were able to reduce the space that electrons have to travel during
calculations. The end result was a $50 million black box designed to operate at
a peak rate of 10 billion calculations per second, 30 times faster than
previous supercomputers.